limma vs DEqMS for Proteomics — When to Use Which (n=3 to n=20+ Comparison)

The Question That Comes Up Every Proteomics Paper

You have a DIA-NN or MaxQuant output. 4,000 quantified proteins. n=3 per group. Reviewer 2 says: "Why did you use t-test instead of limma?" Co-author says: "Why didn't you use DEqMS specifically for proteomics?"

Both are reasonable points. The difference between limma and DEqMS in 2026 proteomics is not large, but it isn't zero, and the right choice depends on whether your data has the metadata DEqMS needs (peptide counts per protein) and whether the small-effect-size proteins matter for your conclusions.

This guide is the practical comparison: when DEqMS's proteomics-specific variance prior actually helps, when it's noise compared to plain limma, and when even an unmoderated Welch's t-test would be fine.

It builds on the general statistical-test selection picture in Statistical Test Selection Guide — t-test, limma, ANOVA.

What Each Tool Actually Does

Plain Welch's t-test

Per protein: compute t-statistic on log2 intensities, two-sided p-value, BH-FDR across all proteins.

• Strength: simple, no assumptions about cross-protein structure
• Weakness: with small n (n=3), the per-protein variance estimate is unstable — a few proteins with accidentally tiny variance get inflated t-statistics → false positives

limma (Smyth, 2004)

Empirical Bayes shrinkage of variance. Borrows information across all proteins to stabilize per-protein variance estimates, then computes moderated t-statistics. Originally designed for microarrays but routinely used for proteomics since 2010.

• Strength: massive improvement over t-test for small n (n=3-5) — typically 20-40% more true positives at the same FDR
• Weakness: doesn't know proteins differ in measurement quality (a protein quantified by 1 peptide is noisier than one with 20 peptides)

DEqMS (Zhu et al., 2020)

Builds on limma but adds a peptide-count-dependent variance prior: proteins with more peptides get smaller residual variance estimates (a real proteomics property). The result is a proteomics-aware moderated t-statistic.

• Strength: better calibration of significance for low-peptide proteins (single-peptide IDs are downweighted appropriately)
• Weakness: requires per-protein peptide count metadata, which DIA-NN and MaxQuant both provide but you have to wire it up

When Each Wins — A Practical Decision Tree

Do you have n ≥ 30 per group?
  → Welch's t-test is fine. Skip limma/DEqMS.

Do you have peptide-count per protein in your output?
  YES → DEqMS is the right default
  NO  → limma is the right default (or compute peptide counts and use DEqMS)

Are your DEPs heavily weighted toward single-peptide proteins?
  YES → DEqMS will materially change conclusions
  NO  → limma and DEqMS give very similar lists

For most modern proteomics outputs (MaxQuant proteinGroups.txt, DIA-NN report.pg_matrix.tsv + report.tsv) you DO have peptide count, so DEqMS is the recommended default in 2026.

Side-By-Side Code

Setup

library(limma)
library(DEqMS)
library(readr)

# Load MaxQuant proteinGroups.txt (or DIA-NN pg_matrix.tsv with peptide count joined)
pg <- read_tsv("proteinGroups.txt")

# Standard filters
pg <- pg[pg$Reverse != "+" & pg$`Potential contaminant` != "+" & pg$`Only identified by site` != "+", ]

# Identify intensity columns (LFQ or iBAQ)
lfq_cols <- grep("^LFQ intensity ", colnames(pg), value = TRUE)
mat <- as.matrix(pg[, lfq_cols])
rownames(mat) <- pg$`Majority protein IDs`

# Replace zeros with NA, log2 transform
mat[mat == 0] <- NA
mat <- log2(mat)

# Group assignments
group <- factor(c("Treat","Treat","Treat","Treat","Ctrl","Ctrl","Ctrl","Ctrl"))
design <- model.matrix(~0 + group)
colnames(design) <- levels(group)
contrast <- makeContrasts(Treat - Ctrl, levels = design)

limma path

fit <- lmFit(mat, design)
fit <- contrasts.fit(fit, contrast)
fit <- eBayes(fit)
limma_results <- topTable(fit, number = Inf, adjust.method = "BH")

DEqMS path (adds 3 lines)

fit <- lmFit(mat, design)
fit <- contrasts.fit(fit, contrast)
fit <- eBayes(fit)

# DEqMS extension — peptide count per protein
fit$count <- pg$`Peptides`     # MaxQuant column. For DIA-NN, count unique stripped peptides per protein
deq <- spectraCounteBayes(fit)
deqms_results <- outputResult(deq, coef_col = 1)

That's the actual delta — three extra lines that link variance modeling to per-protein measurement quality.

Real Comparison on a 4,000-Protein Dataset

A typical cross-species ECM proteomics analysis (n=4 per group, ~4,000 quantified proteins after valid-value filter):

| Method | DEPs (adj.p < 0.05, |log2FC| ≥ 1) | Single-peptide DEPs | |---|---|---| | Welch's t-test | 312 | 87 (28%) | | limma | 416 | 75 (18%) | | DEqMS | 397 | 22 (6%) |

Two patterns:

1. limma > t-test by ~30%: empirical Bayes is doing real work — recovering true positives that t-test misses because of unstable variance at n=4
2. DEqMS < limma in total count but much lower fraction of single-peptide DEPs: DEqMS correctly downweights low-peptide proteins, which is closer to biologically defensible truth

The 53 proteins that limma calls DEP but DEqMS does not are mostly single-peptide hits — exactly the proteins where caution is appropriate.

What if you don't care about single-peptide proteins?

If you filter out single-peptide proteins upfront (a defensible choice for many downstream analyses):

pg_filtered <- pg[pg$Peptides >= 2, ]

After this filter, limma and DEqMS converge — usually within 5% of each other on DEP count. DEqMS still slightly downweights 2-peptide vs 20-peptide, but the practical difference shrinks.

Sample Size Effects

n = 3 per group

• Welch's t-test: highly unstable; many false positives, many missed real signals
• limma: dramatic improvement, the empirical Bayes shrinkage is exactly what small-n needs
• DEqMS: similar to limma, slightly better calibration on low-peptide proteins

Recommendation at n=3: limma is mandatory, DEqMS preferred if peptide counts available.

n = 5-10 per group

• Welch's t-test: usable but suboptimal
• limma: still meaningfully better
• DEqMS: marginal improvement over limma

Recommendation: limma minimum, DEqMS preferred.

n = 15-30 per group

• Welch's t-test: getting close to limma performance
• limma: still slightly better
• DEqMS: nearly indistinguishable from limma

Recommendation: limma still preferred for consistency; DEqMS overhead not worth it.

n ≥ 30 per group

• Welch's t-test: essentially equivalent to moderated tests
• limma / DEqMS: no longer adding value

Recommendation: plain Welch's t-test or even Wilcoxon (non-parametric) is fine. The empirical Bayes is no longer needed.

Common Pitfalls

1. Forgetting log2 transformation

Both limma and DEqMS expect log-transformed intensities. Run them on raw LFQ and the results are nonsense — yet no error message will warn you. Always log2 transform first.

2. NA handling

# Bad — discards entire protein if any sample missing
mat_complete <- mat[complete.cases(mat), ]

# Better — limma handles NAs within group as long as ≥2 per group
# (use Valid-Value filter explicitly per [the cross-species pillar])

See Reproducing Park et al. 2026 for the valid-value filter pattern.

3. Using LFQ when iBAQ is more appropriate (or vice versa)

• LFQ (Label-Free Quantification): MaxQuant's normalized intensity, designed for cross-sample comparison
• iBAQ: intensity per peptide, designed for relative protein abundance within a sample

For most differential expression with MaxQuant, LFQ is the right starting point. DIA-NN's PG.MaxLFQ is the equivalent. Don't run limma/DEqMS on raw Intensity columns.

4. Wrong peptide count column in DEqMS

DEqMS's fit$count expects the number of peptides quantifying each protein. In MaxQuant proteinGroups.txt this is the Peptides column (sometimes Razor + unique peptides is more appropriate). In DIA-NN, count unique stripped peptide sequences per Protein.Group:

# For DIA-NN
library(dplyr)
peptide_counts <- diann_report %>%
  group_by(Protein.Group) %>%
  summarize(count = n_distinct(Stripped.Sequence))
# join into pg_matrix

If you use the wrong column (e.g., Unique peptides when you should use Peptides), DEqMS will run but its variance modeling will be wrong.

5. FDR interpretation

Both limma and DEqMS report adjusted p-values (BH-FDR). At FDR 0.05, expect roughly 5% false discovery in your DEP list. For high-stakes claims (biomarker proposals, drug targets), apply tighter cutoffs (FDR 0.01) and confirm with orthogonal methods.

When Neither Is Enough

Some scenarios where you need to go beyond limma/DEqMS:

• Batch effects: use removeBatchEffect() first, or include batch as a covariate in the limma design matrix
• Repeated measures / paired samples: use duplicateCorrelation() in limma
• Multiple groups with interactions: standard limma design matrix syntax handles this
• Heavy-tailed distributions (some PTM datasets): consider non-parametric (Wilcoxon) or robust regression
• Highly non-uniform missingness: model MNAR explicitly (see the imputation methods post)

What If You Don't Have R?

DEqMS is R-only. Python alternatives:

• statsmodels for plain Welch's t-test + BH
• limma via rpy2 — call R from Python (annoying but works)
• Custom moderated t-test in Python: possible but not standard practice; defaults to plain t-test for most Python users

For proteomics specifically, R is still the lingua franca for differential expression. If your pipeline is Python-heavy, consider switching to R just for the DEP step, then back to Python for downstream.

FAQ

Q: Is DEqMS just better than limma for proteomics, period? For data where peptide counts vary dramatically across proteins (which is most LC-MS data), yes — DEqMS gives better-calibrated significance. For datasets where you've filtered to multi-peptide proteins, the difference is small.

Q: What about ProteoMM, MSstats, ProteinMixed?

• MSstats: more general framework, handles paired designs and protein-level inference from peptide-level data. Heavier but well-validated.
• ProteoMM: handles missing data with mixture models. Niche use cases.
• ProteinMixed: research stage; not for production work yet.

For most proteomics, DEqMS or MSstats are the two serious choices. DEqMS is simpler when you start from a protein-level intensity matrix; MSstats is preferred when you have peptide-level data and complex designs.

Q: Can I use DEqMS for DIA-NN output? Yes. The trick is computing peptide counts per protein from DIA-NN's report.tsv (it's peptide-level) and joining to report.pg_matrix.tsv (protein-level). 10 lines of dplyr or pandas.

Q: limma's voom() — should I use that? voom() is for RNA-seq count data (mean-variance modeling for counts). Proteomics intensities are continuous after log2, not counts. Don't use voom for proteomics. Use lmFit + eBayes directly.

Q: What FDR threshold should I use? 0.05 is conventional. For pilot/discovery work where downstream validation will filter, 0.10 is sometimes used. For biomarker proposals or any claim that goes into a paper without orthogonal validation, 0.01 is safer. Discuss with reviewers in your field.

Q: My DEqMS gives weird variance plots with VarianceBoxplot() — what's that mean? DEqMS expects a monotonic decreasing relationship between peptide count and variance (more peptides → lower variance). If your boxplot doesn't show this trend, something is off — usually a normalization or contaminant issue. Fix upstream, then rerun.

Closing — One-Line Answer

For most 2026 proteomics differential expression with n=3-10, use DEqMS if you have peptide counts (you usually do), use limma if you don't, use Welch's t-test only if n ≥ 30. The extra 3 lines for DEqMS over limma are worth it for almost any LC-MS proteomics dataset.

---

Related posts:

• Statistical Test Selection Guide — t-test, limma, ANOVA
• Reproducing Park et al. 2026 — Cross-Species ECM Proteomics, Three Iterations
• From DIA-NN Output to Paper Draft: AI-Assisted Proteomics Workflow
• Imputing Missing Values in Proteomics — knn vs minDet vs MNAR
• FragPipe vs MaxQuant 2026 Speed Benchmark
• LC-MS/MS Proteomics Complete Workflow Guide 2026

References:

• Smyth, G. K. (2004). Linear models and empirical Bayes for microarray. Statistical Applications in Genetics and Molecular Biology, 3, Article 3.
• Zhu, Y. et al. (2020). DEqMS: a method for accurate variance estimation in proteomics. Molecular & Cellular Proteomics, 19, 1047-1057.
• Ritchie, M. E. et al. (2015). limma powers differential expression analyses. Nucleic Acids Research, 43, e47.
• Choi, M. et al. (2014). MSstats: an R package for statistical analysis of quantitative mass spectrometry. Bioinformatics, 30, 2524-2526.